Design And Optimization Of Face Recognition Algorithm For Conference Room Reservation System

Intelligent conference room reservation system enables the conference room to have remote reservation,control,monitoring and other functions,and effectively solve the problem of low efficiency of traditional manual management.Among them,the use of face recognition technology with high recognition accuracy for subscriber identity authentication can ensure the security and effectiveness of the system.Therefore,facing the conference room reservation system,the existing deep learning face recognition algorithm and network structure are improved to achieve high accuracy and speed in this thesis.Firstly,the existing Cosine Loss face recognition algorithm is studied and analyzed.Aiming at the problem that the algorithm must use high norm eigenvectors to make the network converge and affect the recognition accuracy,the influence of feature norm values on the training process is deeply analyzed.A low Norm Cosine Loss(LNCL)algorithm is proposed,which improves recognition accuracy while guarantees the convergence of the network.In LNCL algorithm,the dynamic filtrating strategy of full connected layer reduces the lower bound of training loss and ensures the convergence of the network;low norm eigenvectors enable the network to be fully trained and improve the recognition accuracy of the network.In addition,in order to optimize the performance of convolution neural network,the parameter configuration of ResNet convolution neural network is improved.By optimizing the global step size of the network and using Elu activation function,the accuracy of face recognition is further improved.On this basis,in order to apply the algorithm in the conference room reservation system,the LNCL algorithm and the optimized ResNet are implemented on the cloud server in this thesis,and the video transmission process in the system is optimized,which meets the real-time requirements of the conference room reservation system.The LNCL algorithm is tested with the face recognition benchmark LFW,YTF and MegaFace.The test results show that the LNCL algorithm achieves 99.78% verification accuracy on LFW,96.80% on YTF and 81.70% on Macaface.Then the LNCL algorithm is validated in the actual conference room reservation system.The actual operation results show that the overall recognition rate of face recognition access control in the conference room reservation system based on LNCL algorithm can reach 29.59 FPS,and the face can be accurately recognized in complex real scenes,including multi-pose,wearing/taking off glasses,changing clothes,backlighting and so on.